torch.from_file is not documented

[ezyang]

I'm not sure this was intentionally put in the torch namespace

cc @svekars @carljparker

[ezyang]

@albanD says this is a useful function but shouldn't be called from_file

[albanD]

I feel like this should be a low level function (we currently use it as such for multiprocess Tensor sharing IIRC).
Maybe we should provide a nice high level API for people that want to do disk-backed Tensor (for very large Tensors or sharing across processes).

[afcruzs]

Maybe we should provide a nice high level API for people that want to do disk-backed Tensor (for very large Tensors or sharing across processes).

@albanD @ezyang is as of today torch.from_file the only way to accomplish this? Is it recommended? Basic use case for me would be to read a fraction of a very large tensor stored in disk (without loading it entirely to memory).

[albanD]

Hey!
This is the only way today I'm afraid but #101446 will add support for this in torch.load!

[afcruzs]

Oh this looks very nice, looking forward to it!

[vadimkantorov]

Though torch.load mmap still deals only with certain packed formats. A bare torch.from_file (maybe with mmap option) would still be nice to have for fostering experimentation with ram/disk offloading

[albanD]

Agree, we should add doc for this!